Abstract: In this paper, the second -order difference cochlear model is used to extract the speech parameters. A kind of speech recognition front-end parameters based on auditory spectrum is obtained. A new "amplitude sum multiplied by frequency difference" distance measure is proposed according to the feature of speech parameters. The recognition algorithm is an improved DTW algorithm that sets two free frames in the beginning of speech segments and has the trace slope between 1/2 and 2. Under the recognition condition of small vocabulary or digits vocabulary and speaker independence, computer simulation shows that the algorithm attains an recognition accuracy of at least 98 percent, and it has the quite good robustness as well.

Key words: speech recognition, second-order difference cochlear model, auditory spectrum based speech parameter